Teaching LLMs to Speak Spotify — Yves Raimond & Jacqueline Wood, Spotify

Why watch

Spotify engineers explain how they built their LLM-native recommendation system using the Large Taste Model. Jacqueline Wood details the NEO four-stage training method: adding Semantic IDs as new tokens to open-weight LLMs like Qwen while keeping the backbone frozen (avoiding the language ability loss from continued pre-training). The system achieves 98% valid semantic IDs without constrained decoding, and grounding LLM judges with user profiles improved human agreement by 91% on ambiguous queries. Covers generative personalization features including steerable DJ, prompted playlists, and editable taste profiles.